LLVM 12 regression (GlobalISel): AArch64 backend generates incorrect code for ashr <i8 x 16>

[Amanieu]

Bugzilla Link	49904
Resolution	FIXED
Resolved on	May 10, 2021 17:48
Version	trunk
OS	Linux
Blocks	#48246 #48661
CC	@aemerson,@Arnaud-de-Grandmaison-ARM,@CryZe,@fhahn,@ornata,@nikic,@smithp35,@tstellar
Fixed by commit(s)	40e75ca 24535af

Extended Description

IR:

define <16 x i8> @​simd_ashr(<16 x i8> %a, <16 x i8> %b) {
start:
%0 = ashr <16 x i8> %a, %b
ret <16 x i8> %0
}

llc -march=aarch64 -global-isel=1 -O0 test.ll

simd_ashr: // @​simd_ashr
// %bb.0: // %start
neg v1.8h, v1.8h
sshl v0.16b, v0.16b, v1.16b
ret

The neg instruction should be using .16b instead of .8h.

[Amanieu]

assigned to @tstellar

[aemerson]

Thanks! Fixed in 40e75ca

[nikic]

Reopening for LLVM 12 backport (or 12.0.1, depending on where things stand).

[fhahn]

Thanks! Fixed in 40e75ca

Awesome, thanks Amara!

[tstellar]

Hi Tim,

What is your opinion on backporting this?

https://reviews.llvm.org/rG40e75cafc0fef365b5580a9c09595ac475db0c19

[TNorthover]

Looks very simple to me, so I say yes.

[tstellar]

Merged: 24535af